Apparatus and method for modifying ground truth for checking accuracy of machine learning model

1 . A method for modifying ground truth of image data for testing accuracy of an object detection machine learning model, the method comprising:

obtaining virtual image data of an image capturing area from a camera point of view;

obtaining ground truth data regarding virtual objects in the image capturing area;

determining a total size of a virtual object in the virtual image data by:

individually rendering the virtual object in full; and

calculating an area of the individually-rendered virtual object;

calculating a size of a visible area of the virtual object in the virtual image data;

determining a visibility rate of the virtual object by dividing the size of the visible area by the total size of the virtual object;

determining whether the visibility rate of the virtual object is at or above 0% and below a first predetermined threshold, at or above the first predetermined threshold and below a second predetermined threshold higher than the first predetermined threshold, and at or above the second predetermined threshold higher than the first predetermined threshold; and

based on the visibility rate being at or above 0% and below the first predetermined threshold, omitting the virtual object from the ground truth data;

based on the visibility rate being at or above the first predetermined threshold and below a second predetermined threshold higher than the first predetermined threshold, modifying the ground truth data for the virtual object by the determined visibility rate; and

based on the visibility rate being at or above the second predetermined threshold higher than the first predetermined threshold, not modifying the ground data for the virtual object by the determined visibility rate.

2 . The method of claim 1 , wherein the determining the total size of the virtual object comprises:

obtaining a ground truth object segmentation for a full scene of the image capturing area; and

calculating an area of the virtual object in the ground truth object segmentation.

3 . The method of claim 1 , wherein the calculating the size of the visible area of the virtual object in the virtual image data comprises determining a number of pixels included in the visible area or determining an area of a polygon that bounds the visible area.

4 . The method of claim 1 , wherein the obtaining the image data of the image capturing area comprises obtaining the image data generated by a camera sensor model of an autonomous driving simulator.

5 . The method for modifying ground truth of image data of claim 1 , wherein

modifying the ground truth data for the virtual object with the determined visibility rate comprises using an intersection over union metric.

6 . The method for modifying ground truth of image data of claim 1 , wherein

modifying the ground truth data for the virtual object with the determined visibility rate comprises using a mean average precision metric.

7 . An apparatus for modifying ground truth of image data for testing accuracy of an object detection machine learning model, the apparatus comprising:

memory storing instructions, and

at least one processor configured to execute the instructions to:

obtain virtual image data of an image capturing area from a camera point of view;

obtain ground truth data regarding virtual objects in the image capturing area;

determine a total size of a virtual object in the virtual image data by:

individually rendering the virtual object in full; and

calculating an area of the individually-rendered virtual object;

calculate a size of a visible area of the virtual object in the virtual image data;

determine a visibility rate of the virtual object by dividing the size of the visible area by the total size of the virtual object;

determine whether the visibility rate of the virtual object is at or above 0% and below a first predetermined threshold, at or above the first predetermined threshold and below a second predetermined threshold higher than the first predetermined threshold, and at or above the second predetermined threshold higher than the first predetermined threshold; and

based on the visibility rate being at or above 0% and below the first predetermined threshold, omit the virtual object from the ground truth data;

based on the visibility rate being at or above the first predetermined threshold and below a second predetermined threshold higher than the first predetermined threshold, modify the ground truth data for the virtual object by the determined visibility rate; and

based on the visibility rate being at or above the second predetermined threshold higher than the first predetermined threshold, not modify the ground data for the virtual object by the determined visibility.

8 . The apparatus of claim 7 , wherein the at least one processor is further configured to execute the instructions to:

obtain a ground truth object segmentation for a full scene of the image capturing area; and

calculate, as the total size of the virtual object, an area of the virtual object in the ground truth object segmentation.

9 . The apparatus of claim 7 , wherein the at least one processor is further configured to execute the instructions to calculate the size of the visible area of the virtual object in the virtual image data by determining a number of pixels included in the visible area or determining an area of a polygon that bounds the visible area.

10 . The apparatus of claim 7 , wherein the at least one processor is further configured to execute the instructions to obtain the image data generated by a camera sensor model of an autonomous driving simulator.

11 . A non-transitory computer-readable recording medium having recorded thereon instructions executable by at least one processor to perform a method for modifying ground truth of image data for testing accuracy of an object detection machine learning model, the method comprising:

obtaining virtual image data of an image capturing area from a camera point of view;

obtaining ground truth data regarding virtual objects in the image capturing area;

determining a total size of a virtual object in the virtual image data by:

individually rendering the virtual object in full; and

calculating an area of the individually-rendered virtual object;

calculating a size of a visible area of the virtual object in the virtual image data;

determining a visibility rate of the virtual object by dividing the size of the visible area by the total size of the virtual object;

determining whether the visibility rate of the virtual object is at or above 0% and below a first predetermined threshold, at or above the first predetermined threshold and below a second predetermined threshold higher than the first predetermined threshold, and at or above the second predetermined threshold higher than the first predetermined threshold; and

based on the visibility rate being at or above 0% and below the first predetermined threshold, omitting the virtual object from the ground truth data;

based on the visibility rate being at or above the first predetermined threshold and below a second predetermined threshold higher than the first predetermined threshold, modifying the ground truth data for the virtual object by the determined visibility rate; and

based on the visibility rate being at or above the second predetermined threshold higher than the first predetermined threshold, not modifying the ground data for the virtual object by the determined visibility rate.

12 . The non-transitory computer-readable recording medium of claim 11 , wherein the determining the total size of the virtual object comprises:

obtaining a ground truth object segmentation for a full scene of the image capturing area; and

calculating an area of the virtual object in the ground truth object segmentation.

13 . The non-transitory computer-readable recording medium of claim 11 , wherein the calculating the size of the visible area of the virtual object in the virtual image data comprises determining a number of pixels included in the visible area or determining an area of a polygon that bounds the visible area.

14 . The non-transitory computer-readable recording medium of claim 11 , wherein the obtaining the image data of the image capturing area comprises obtaining the image data generated by a camera sensor model of an autonomous driving simulator.